Object Detection Research Based On Multi-level Feature Pyramid Network

The thought of object detection is applied in the field of computer vision,such as face recognition,license-plate detection and instance segmentation.As the development of object detection methods,there are also some shortcomings.Such as insufficient feature fusion leading to low detection performance,the less semantic information of high-level small objects leading to a decrease in detection performance.The research of multi-level feature fusion and contextual detail information have important research significance for object detection.This dissertation focuses on the multi-scale feature pyramid network and the specific content is as follows.This dissertation uses a multi-scale feature pyramid network to study the effect of detection.Introducing feature maps of different resolutions to detect the features of objects,so that the accuracy of its detection is better than that of single object detection.This dissertation improves the Multi-Branch Multi-Level Semantic and Context feature for object detection-DMLSCF.The network realizes the fusion of different feature layers,the enhancement of contextual details,and the detection of objects with multi-resolution feature maps.The experimental results show that using the PASCAL VOC data set,the test result of the network when the input size is 512×512is 84.3%,1.0% better than the MLFCPNet network,and 1.4% better than the current best network.Using the COCO data set,when the input size is 512×512,the test result of the network is 38.2% better than MLFCPNet by 1.1%,and 1.5% better than the current optimal model under the same backbone.This dissertation proposes a method to use the feature fusion module for the adjacent three-layer features of the backbone.This method can effectively solve the problem of insufficient feature fusion.The backbone uses this method to enhance its feature fusion ability.Experimental results show that the feature fusion module is very effective for DMLSCF network,and the experimental results on the PASCAL VOC data set is 3.5% higher than without adding the module.This dissertation proposes a method that uses contextual detail features for object detection,and fuses semantic information with contextual detail features to enhance feature information.The experimental results show that the context network is effective in improving the DMLSCF network,and the experimental results on the PASCAL VOC data set is 0.6% higher than without adding the network.